Chapter 5

DIVERSITY COMBINING

This material to appear as part of the book Space-Time Methods, vol. 1:

Space-Time Processing
c
!2012 by Sebastien Roy. All rights reserved.

1. The diversity concept

The term “diversity" refers to the availability at the receiver of multiple
copies of the desired signal, each of which being affected by different channel
characteristics. This allows the implementation at the receiver of a combiner,
which generally consists of a weight-and-sum structure where the weights are
chosen to obtain the best possible estimate of the desired signal.
While diversity is mostly useful to combat fading due to multipath, it can
also provide a gain against white noise and interference. The different types of
diversity include:

Time diversity — the same data is transmitted repeatedly several times; a

buffer at the receiver can be used to store and combine the multiple copies.

Frequency diversity — the data is transmitted simultaneously in multiple

frequency bands which may have completely or partially uncorrelated fad-
ing characteristics.

Space diversity — the receiver is equipped with an antenna array made

up of N antenna elements. In a multipath fading environment, sufficient
spacing can be introduced between the elements to decorrelate the fading
envelopes.

309
310 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

Impulse diversity (also known as multipath diversity)— in CDMA or other

systems where the band is sufficiently large to resolve the individual im-
pulses in the impulse response, each individual impulse constitutes a di-
versity channel with its own fading characteristics. They can be weighted
and summed by a RAKE combiner to achieve the same benefits as in time,
frequency or space diversity [Proakis, 2000, ch. 14].

It follows that analyses of simple diversity schemes remain valid regardless

of the diversity source (space, time, frequency or multiple rays with RAKE
combining).

2. Scanning diversity
Scanning diversity is possibly the simplest form of linear processing since
only a single RF receiver front-end needs to be implemented. Furthermore, it
does not require estimation of the SNR on all branches simultaneously (like
selection diversity, see below) which would entail additional implementation
complexity. Two strategies are in common usage:

1. Switch and examine: The single receiver is connected to one of the an-
tenna elements on the array via a switch, and it stays connected to the same
element until the SNR drops below a predefined threshold. At that point,
it is switched to other antenna elements in succession until one is found
which yields a signal power above the threshold. The disadvantage of this
method is that, if no element has a SNR above the threshold, a large amount
of useless switching activity results, causing noise and a waste of power.
2. Switch and stay: The switch-and-stay strategy solves the switching prob-
lem which occurs with switch and examine. Here, a new branch is selected
when the SNR on the current branch crosses the threshold from high to low.
However, it stays on the newly selected branch even it its SNR is lower than
the threshold until here also, the SNR crosses the threshold from high-to-
low. While no rapid switching activity is generated, this strategy can cause
the system to wait on a poor SNR branch until it rises above threshold and
falls again, even if better branches might be available.

3. Selection diversity
This minimalist diversity strategy consists of selecting only one branch out
of the array (again through switching means), the one with the highest instanta-
neous SNR. Like scanning diversity, only a single complete receiver is required
(in principle). However, partial reception circuitry might be necessary on all
branches to permit SNR monitoring.
If all branches are uncorrelated and have identical average powers, the anal-
ysis of the performance of selection diversity is relatively simple and is based
Diversity Combining 311

on order statistics. Indeed, given N random variables x1 , . . . , xN which are

independent and identically distributed (i.i.d.), we can map these to a set of
ordered r.v.’s y1 , . . . , yN such that y1 is always equal to the largest value in the
set x1 , . . . , xN , y2 always equal to the second largest, etc.
Formally, we have
!
i = 1, . . . , N,
y i = xn i , (5.1)
ni "= nj , ∀i "= j,
and

y1 ≤ y2 · · · ≤ yN . (5.2)
The density of yk is given by

N!
fyk (y) = Fxk−1 (y) [1 − Fx (y)]n−k fx (y), (5.3)
(k − 1)!(N − k)!
where Fx (x) is the cumulative distribution function of the i.i.d. r.v.’s xi and
fx (x) is their density. Hence, the distribution of the SNR γ of an N -element
selection diversity array corresponds simply to the distribution of yN , i.e.

fγ (γ) = N FxN −1 (γ)fx (γ), (5.4)

where fx (x) and Fx (x) correspond respectively to the PDF and CDF of the
branch SNR.
Another, possibly simpler, approach to this problem stems from the obser-
vation that the CDF of the selection diversity SNR, i.e. the probability that the
array SNR γ is smaller than some value x, is equal to the probability that all
the branch SNRs are smaller than x. Hence,

Fγ (γ) = Fx (γ)N . (5.5)

It can be verified that deriving the above will yield (5.4).
If the SNRs at the branches do not follow identical distributions but remain
uncorrelated, order statistics no longer apply. However, the CDF approach of
(5.5) remains valid and we have
N
"
Fγ (γ) = Fxn (γ), (5.6)
n=1

where Fxn (x) is the CDF of the nth branch SNR. Likewise, if the branch SNRs
are correlated, we have

Fγ (γ) = Fx1 ,x2 ,··· ,xN (γ, · · · , γ), (5.7)

312 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

where Fx1 ,x2 ,··· ,xN (x1 , x2 , · · · , xN ) is the joint CDF of the branch SNRs. How-
ever, this joint CDF does not have a known closed form for any N ≥ 2 for all
common fading distributions.

Rayleigh fading environments

For uncorrelated, identically distributed Rayleigh fading, the PDF of the
branch SNR is a scaled, 2 complex degrees-of-freedom chi-square variate, i.e.
1 − γx¯
fx (x) = e c u(x), (5.8)
γ̄c
and the corresponding CDF is
x
Fx (x) = 1 − e− γ̄c . (5.9)
Therefore, we have
N γ γ
fγ (γ) = (1 − e− γ̄c )N −1 e− γ¯c . (5.10)
γ̄c
The basic probability of error for noncoherent DPSK is given by
1
PbDP SK (γ) = e−γ . (5.11)
2
Hence, the probability of error for noncoherent DPSK of a selection diver-
sity combiner operating in Rayleigh fading is given by

# ∞
P2DP SK = fγ (γ)PbDP SK (γ)dγ
0
# ∞
N
! "
1
− γ̄γ N −1 −γ γ¯c +1
= (1 − e c ) e dγ, (5.12)
2γ̄c 0
Although the above integral is not readily available in tables, it is reduced
in Appendix 5.A to yield
$ %
DP SK 1 1 + γ̄c , N + 1
P2 = Γ . (5.13)
2 1 + γ̄c + N

Practical issues
Classical performance results can be found in [Jakes, 1974], [Pierce, 1958].
Classical theoretical results on selection diversity assume (1) that SNR mea-
surements are noise-free and (2) that based on these measurements, the switch
to the best branch is instantaneous. While this is obviously analytically conve-
nient and useful for many purposes — including comparing diversity strategies
Diversity Combining 313

— it does not yield a realistic performance assessment. Indeed, it is difficult in

practice to extract only the desired signal power (thus rejecting noise and inter-
ference) and even if this is achieved, some degree of noise invariably remains.
Some authors have looked at the impact of switching constraints [Ritcey
and Azizo, 1998], [Barnard and Pauw, 1989] which include physical switch-
ing delays and system-imposed dwell times to avoid information loss during
switching.
It was recognized in [Neasmith and Beaulieu, 1998] that typical power
measurement devices in practical selection combiners would yield signal-plus-
noise power instead of signal-to-noise ratio. The performance analysis therein
shows that combiners based on signal-plus-noise measurements perform bet-
ter than combiners based on the signal-to-noise ratio measure. However, this
assumes that (1) the system has full receiver chains on each branch, (2) it is
capable of picking the best branch on a symbol-by-symbol basis and (3) it se-
lects the said branch based on S+N measurements for the symbol to be detected.
Hence, a branch has more chance of being selected when the noise sample adds
up constructively with the useful signal. However, and while such a strategy
may sometimes make sense (since it is numerically simpler than maximal-ratio
combining for example), it is impossible to implement with a single receiver
chain and in that sense defeats the purpose of using selection diversity.
In the same vein, improved selection diversity strategies based on branch
log-likelihood ratios and yielding even superior performance were presented
in [Kim and Kim, 2001]. However, these make the same assumptions as in
[Neasmith and Beaulieu, 1998] and are therefore not cost-effective SDC im-
plementations from an RF / hardware point-of-view.

A practical selection diversity implementation

This section examines the use of Schottky diodes to measure signal-plus-
noise power in each branch and the use of PIN diodes as switching means to
select the appropriate branch to pass on to a single receiver chain. The architec-
tural elements discussed are by no means novel and should seem “natural” to a
practitionner of the art, they are presented as low-cost methods of implement-
ing selection diversity and accompanied by a precise performance analysis.
To take full advantage of the simplicity inherent to selection diversity, the
receiver circuit should be simplified and the component count minimized. To
achieve this, RF switching means are employed in conjunction with a sin-
gle downconversion / receiver chain. Only very inexpensive components (e.g.
Schottky and PIN diodes, amplifiers, etc.) need to be replicated in each branch.
This approach also implies that the cost of the receiver is relatively insensitive
to the number of branches.
A block diagram of the proposed RF front-end is shown in Figure 5.1. Di-
rectional couplers are used at the outputs of the LNAs in each branch to divert
314 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

a tiny fraction of the incident power to envelope detectors for the purpose of
monitoring the total signal power. Based on these power measurements, branch
selection can be performed via a single pole multiple throw (SPMT) switch.
For low-cost and other practical considerations, it makes sense to construct
such switches out of PIN (positive-insulator-negative) diodes which are char-
acterized by their fast switching time and moderate insertion loss (less than 1
dB). See [Wang et al., 2001] for a detailed description of a PIN diode based
rotary switch.
It is noteworthy that the switching transient has a duration proportional to
the reciprocal of the IF bandwidth and we can expect a few of the subsequent
symbols to be corrupted by the transition from one branch to another. In fact,
the relatively low transition time of the switch induces an almost instantaneous
phase/amplitude shift which, when propagating through the remaining receiver
stages expresses the impulse response of these cascaded components. Taking
this into account, switching can be restriced to non-information-bearing peri-
ods, i.e. during packet headers, tail bits and / or guard times.
Figure 5.2 details an envelope detector built around a low-cost Schottky
diode. The small signal approximation of the diode current for a given voltage
v(t) applied at the diode is [Pozar, 1998]:

v2 #
Id (v) = I0 + vGd + G (5.14)
2 d

where Gd is the dynamic conductance of the diode and I0 is the bias current.
Let us define a modulated signal

vin (t) = Re{α(t)ejωc t } (5.15)

= I(t) cos ωc t − Q(t) sin ωc t (5.16)

where α(t) is the signal envelope, I(t) = Re{α(t)} and Q(t) = Im{α(t)} are
the inphase and quadrature components of the baseband signal, respectively,
and Re{·} and Im{·} denote the real and imaginary parts of their arguments.
Substituting (3) into (1) yields the output current. Taking into account the fact
that the acquisition chain necessary to sample the output of the envelope de-
tector (e.g. differential amplifier, analog-to-digital converter, etc.) necessarily
includes some form of low-pass filtering, terms at ωc t and 2ωc t are rejected
and the baseband current is given by ωc t and 2ωc t) is:

G#d & 2
I (t) + Q2 (t)
'
IBB (t) = I0 + (5.17)
4
G#d
= I0 + |α(t)|2 . (5.18)
4
Diversity Combining 315

The voltage induced by this current across the video resistance of the detector
(also the output voltage of the detector) is:

VBB (t) = K + β|α(t)|2 (5.19)

where β (the voltage sensitivity) and K (related to the bias current) are constant
terms.
The small signal approximation typically remains valid for input powers
up to -20 dBm, i.e. the detector is in its sqare-law region; beyond this point,
the approximation no longer holds and a higher order model is called for. The
tangential sensitivity of the detector typically is near -55 dBm for a video band-
with of 1MHz. Below that threshold, the output of the diode is submerged by
thermal noise and no power measurement is available. In order to boost the
sensitivity of the envelope detectors, LNAs are placed ahead of the envelope
detector in Figure 1 (thus requiring replication accross all branches). However,
if the receiver is extremely sensitive, the possibility remains that the envelope
detectors will be less sensitive than the receiver (depending on the LNA gain).
Nonetheless, the practical impact is minimal since (1) it would only affect per-
formance (by making incorrect branch selections) when all branches are in a
deep fade, and (2) the power measurements would be unreliable in any case
since they are performed on the total signal-plus-noise power.
Another approach proposed to achieve selection diversity would be to first
filter and amplify the signal from each antenna and then apply this pre-processed
signal to the switch as shown in Figure 5.3. With this strategy, the power
monitoring for the antennas is performed sequentially and must be completed
during guard times since switching operations are now required. Therefore,
diversity branches can be scanned in sequence and the best branch selected
before information-bearing transmission is resumed. One advantage here is a
further simplification of the receiver and a higher signal monitoring sensitivity.
Indeed, the signal gain at the output of the downconverter and IF amplifiers is
higher and the signal presented at the input of the detector is more likely to
exceed its tangential sensitivity point. The downside, however, is the longer
guard times required due to the additional switching.
Let sn [k] be the useful signal on the nth branch and νn [k] be the noise sample
on the nth branch, both quantities being sampled at the output of the matched
filter at sampling instant t = kT . A single Schottky signal power measurement
on antenna n is given by

an = |αn |2 , αn = sn + νn , (5.20)

where an is effectively the signal-plus-noise power and where the dependence

on time t is implicit.
316 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

Therefore, the instantaneous signal-to-noise ratio of branch n is

|sn |2 |αn − νn |2
γn = = , (5.21)
σν2 σν2

where σν2 is the noise power .

It follows that the signal-to-noise ratio at the output of the combiner can be
expressed
N N N N
( " ( |αn − νn |2 "
γ= γn u(an − ak ) = u(an − ak ), (5.22)
σν2
n=1 k=1 n=1 k=1
k != n k != n

where !
0, x < 0,
u(x) = (5.23)
1, x ≥ 0,
is the Heaviside step function.
If N = 2, we have

γ = u(a1 − a2 )γ1 + u(a2 − a1 )γ2 . (5.24)

It follows that the distribution of γ conditionned on a1 and a2 is given by

fγ|a1 ,a2 (γ|a1 , a2 ) = u(a1 −a2 )fγ1 |a1 (γ|a1 )+u(a2 −a1 )fγ2 |a2 (γ|a2 ). (5.25)

Hence, we have
# ∞# ∞
fγ (γ) = fγ|a1 ,a2 (γ|a1 , a2 )fa1 (a1 )fa2 (a2 )da1 da2 . (5.26)
0 0

Given the vector of random variables xTn = [an , νn ], its correlation matrix
is given by $ 2
σs + σν2 σν2
%
Rxx = , (5.27)
σν2 σν2
where σs2 is the average useful signal power.
It follows, by virtue of theorem 1.2.11 in [Muirhead, 1982] that the density
2
of νn conditionned on an is Gaussian, with a mean of αn σ2σ+σν
2 and a variance
) 2 * s ν

of σν2 σ2σ+σs
2 .
ν s
From (5.21), we have
# $2
σs2 2 +σ 2
σ 2 + σ 2 − an2 an γn σ 2 σs
+ ,
−γn ν
2 +σ 2
fγn |an (γn |an ) = s 2 ν e σν σs ν
0 F1 1; 2 2 s e σs2
,
σs σν σa
(5.28)
Diversity Combining 317

which is a scaled 2-degrees-of-freedom non-central chi-square distribution.

It is useful to observe that, in substituting (5.25) into (5.26), both terms of
(5.25) will yield the same result after integration, provided that fγ1 |a1 (γ1 |a1 ) =
fγ2 |a2 (γ2 |a2 ) (a1 and a2 simply exchange roles in the two integrals). There-
fore, it suffices to solve the alternate form
$2
2
#
∞# ∞ σs
σ2 + σ2 b1 γn σ 2
# b1 + ,
− 2 2 +σ 2
fγ (γ) = 2 s 2 ν u(b1 − b2 )e σν σs ν
0 F1 1; 2 2s ×
σs 0 0 σν σa
# $
2 +σ 2
σs
−γ ν
σs2
e fb1 (b1 )fb2 (b2 )db1 db2 , (5.29)

where we can remove the step function by using appropriate integration bounds
to yield
$2
2
#
∞ # b1 − b1 σs
σ2 + σ2 b1 γn σ 2
# + ,
2 2 +σ 2
fγ (γ) = 2 s 2 ν e σν σs n
0 F1 1; 2 2s ×
σs 0 0 σν σa
# $
2 +σ 2
σs
−γ ν
σs2
e fb1 (b1 )fb2 (b2 )db2 db1 . (5.30)

If the fading is Rayleigh with variance σs2 on all branches, and the fading
process is uncorrelated accross the array, we have
2
e−x/σa
fa1 (x) = fa2 (x) = , (5.31)
σa2

where σa2 = σs2 + σν2 .

Substituting in (5.29) and integrating over b2 , we get
4 2 +σ 2
# $ # $
∞+ b1
, b1 σs σs
2 ν
#
− − 2 1+ 2 σ2 −γ 2
fγ (γ) = 1−e e 2
σa σa σν a σa2 e σs
×
σs2
0
b1 γσs2
+ ,
0 F1 1, 2 2 db1 , (5.32)
σν σa

which can, in turn. be integrated term-by-term by virtue of Lemma 1.3.3 in

[Muirhead, 1982] to yield
# $
σs2 +σ 2
2σa4 σν2 σa2 σs2
+ , ν
−γ 2
σs
fγ (γ) = 1 F1 1; 1; γ 2 2 e −
σν2 σa2 + σs4 σν σa + σs4
# $
σs2 +σ 2
2σν2 σa2 σa2 σs2
+ , ν
−γ 2
σs
2 2 4 1 F1 1; 1; γ 2 2 4
e , (5.33)
2σν σa + σs σν σa + σs
318 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

where 1 F1 (a; b; x) is the confluent hypergeometric function. The latter has the
property1
x
1 F1 (q; q; x) = e , (5.34)
which, applied to (5.32) yields
# $
σν2 σ4
2σa4 σν2 −γ a
2 σ 2 σ 2 +σ 6
σs
fγ (γ) = 2 2 2 4
e ν a s
σs (σν σa + σs )
# $
2σν 2 σ4
2σa4 σν2 −γ a
2 σ 2 σ 2 +σ 6
2σs
− 2 e ν a s . (5.35)
σs (2σν2 σa2 + σs4 )
If N = 3, we can still cast the problem in the form of (5.29):
2
σ 2 ∞ b1 b1 − σb12 σσs2 b1 γn σ 2
# # # + ,
fγ (γ) = 3 a2 e ν a 0 F1 1; 2 2s ×
σs 0 0 0 σν σa
# $
2 +σ 2
σs
−γ ν
σs2
e fb1 (b1 )fb2 (b2 )fb3 (b3 )db2 db1 . (5.36)
Substituting the Rayleigh fading densities, we find that integration over b2
and b3 yields
# ∞+ b
,2 b1 # σs2 $ #
2 +σ 2
σs ν
$
3 − 12 − 2 1+ 2 −γ 2
fγ (γ) = 1 − e σa e σa σν
σa2 e σs
×
σs2 0
b1 γσs2
+ ,
0 F1 1, 2 2 db1 , (5.37)
σν σa
which can be integrated term-by-term in the same fashion to give
# $
σν2 σ4
3σa4 σν2 −γ a
2 σ 2 σ 2 +σ 6
σs
fγ (γ) = 2 2 2 4
e ν a s −
σs (σν σa + σs )
# $
2σν2 σ4
6σa4 σν2 −γ a
2 σ 2 σ 2 +σ 6
2σs
2 2 2 4
e ν a s +
σs (2σν σa + σs )
# $
3σν2 σ4
3σa4 σν2 −γ a
2 σ 2 σ 2 +σ 6
3σs
e ν a s . (5.38)
σs2 (3σν2 σa2 + σs4 )
Likewise, it is straightforward to show (applying the same term-by-term
integration technique) that the general solution for arbitrary N is
# $
N −1 + (n+1)σν 2 σ4
4σ2
, a
( N −1 σ a ν
−γ 2 σ 2 σ 2 +σ 6
fγ (γ) = N (−1)n 2 e (n+1)σs ν a s .
n σ s ((n + 1)σ 2 σ2 + σ4)
ν a s
n=0
(5.39)

1 This property is obvious when expanding F into its power series representation; however, see also
1 1
[Prudnikov et al., 1990, 7.11.1-4].
Diversity Combining 319

It is simple to show that the average SNR at the output of a classical selection
diversity combiner is
N −1 +
(−1)n
,
( N −1
γ̄ = N γ̄c . (5.40)
n (n + 1)2
n=0

Likewise, simple term-by-term integration of (5.39) gives the average SNR

of the proposed selection diversity architecture:
N −1 +
(n + 1)σs2 σν2 σa2 + σs6
,
( N −1
γ̄ = N . (5.41)
n (n + 1)2 σν2 σa4
n=0

It is noteworthy that the general result (5.39) for the density of the output
SNR can be expressed as follows:
N −1 +
- /0
N − 1 (−1)n σs2 σ 2 νσa2 (n + 1) + σs4
, .
(
fγ (γ) = N gγ γ, ,
n 1+n 0 (n + 1)σν2 σa4
n=0
(5.42)
where ) *N −1 γ0
1 γ̄c
γ
e− γ̄c
gγ0 (γ, γ̄c ) = , (5.43)
γ̄c Γ(N )
which is a 2 degrees-of-freedom scaled central chi-square distribution and it
corresponds to the ideal SNR at the output of a single branch receiver operating
in a Rayleigh fading environment with an average SNR of γ̄c .
It follows that the output SNR distribution (5.42) can be assimilated to a lin-
ear mixture of the output SNRs of N − 1 ideal single-branch receivers. Like-
wise, the bit error probability for any modulation scheme is the same linear
mixture of bit error probabilities for the corresponding ideal single branch re-
ceivers, i.e.
N −1 +
- . /0
N − 1 (−1)n σs2 σν2 σa2 (n + 1) + σs4
,
(SDC2)
(
P2 =N P2 ,
n 1+n (n + 1)σν2 σa4
n=0
(5.44)
where P2 (γ̄c ) is the corresponding bit error rate for a single-branch receiver
averaged over Rayleigh fading with a mean SNR of γ̄c .
(M RC)
This is convenient because bit error probabilities P2 in Rayleigh fad-
ing are well-known for virtually any modulation scheme and can be directly
substituted in (5.44). For example, it is known that [Proakis, 2000]
+ 1 ,
(QP SK) 1 γ̄c
P2 (γ̄c ) = 1− , (5.45)
2 1 + γ̄c
320 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

which, substituted in (5.44), yields

N −1 +
N ( N − 1 (−1)n
,
(SDC2,QP SK)
P2 = (5.46)
2 n=0 n 1+n
- 2 0
σs2 (σν2 σa2 (n + 1) + σs4 )
1− .
σs2 (2σν2 σa2 (n + 1) + σs4 ) + (n + 1)σν4 σa2

Similar expressions can be found for other modulation schemes using the
same approach. Also, it is easy to show (through integration-by-parts on a
term-by-term basis) that the bit error probability for classical selection diver-
sity combining is
N −1 +
2
N ( N − 1 (−1)n
,
(SDC1,QP SK) 2γ̄c
P2 =√ . (5.47)
2 n=0 n n + 1 2(n + 1) + γ̄c

Figure 5.4 compares the PDF of the classical (SDC1) and the proposed
(SDC2) selection diversity architectures for 2, 4 and 10 antenna elements when
operating in Rayleigh fading. It can be seen that the shape of the distribution
is approximately the same for SDC1 and SDC2, with SDC2 suffering a slight
offset to the left. As could be expected, the said offset grows with the number
of antennas.
Figure 5.5 compares the average output SNR of SDC1 and SDC2 for 2, 4,
10 and 25 antenna elements. Three main observations can be drawn: (1) the
curves for the two architectures are relatively close; (2) the number of antennas
affects the slope (thanks to the array gain versus multipath fading); (3) at low
SNRs, SDC2 performs worse than the next smaller array with SDC1.
Figure 5.6 illustrates bit error probabilities averaged over Rayleigh fading
for a QPSK-modulated signal. It can be seen that the BER of SDC2 converges
towards the BER of SDC1 with increasing SNR. With decreasing SNR, the
BER of SDC2 converges towards that of a single-antenna receiver (since it
becomes equivalent to random branch selection as the decisions become in-
creasingly unreliable).
Diversity Combining 321

4. Maximal-ratio combining
Maximum-Ratio Combining (MRC) is a simple and effective combining
scheme for adaptive antenna arrays to combat noise, fading and, to a certain
degree, co-channel interference. However, it requires some form of estima-
tion of the spatial signature (i.e. the channel gain and phase at each antenna
element) of the desired signal accross the array.
MRC is equivalent to optimal MMSE combining when only white noise
is present. It follows that in interference-free environments, it is the optimal
linear combiner. Each branch is simply weighted in proportion to its SNR and
all branches are co-phased and summed. It is, in effect, a spatial matched-
filter. The main difficulty in implementing the maximal ratio concept in its
truest form is the need for perfect knowledge of the vector channel’s amplitudes
and phases. This is, however, a somewhat less severe requirement than with
optimal combining where both the array covariance matrix and the desired
channel must be estimated. Aside from being less computationally intensive
than optimum combining (which requires covariance matrix estimation, matrix
inversion and vector-matrix multiplication), MRC lends itself rather well to
analysis in many application contexts.
If the received signal is affected by white Gaussian noise or white Gaussian
noise and flat fading (i.e. there is no significant man-made interference such
as jamming signals, impulse noise, etc.), the optimal linear weight-and-sum
structure is the maximal-ratio combiner. In MRC, the weight vector in the
complex baseband equivalent domain is simply equal to the complex conjugate
of the desired user vector channel or signature c0 , i.e.

w = c∗0 . (5.48)

Furthermore, this principle is applicable directly to the first three types of

diversity enumerated in section 1 and it is well-known that RAKE combining,
when the paths are weighted proportionally to their SNR, is equivalent to MRC
performance-wise [Proakis, 2000].
It will be assumed herein that the receiver is equipped with an N -element
antenna array and exploits spatial diversity. It follows that the signature c0 has
N elements. The principles outlined and the results obtained, however, are
equally applicable to any of the four diversity types.

Ideal MRC performance

In complex baseband notation [Proakis, 2000], the input to an N -branch
combiner is given by

x(t) = c0 (t)s0 (t) + n(t), (5.49)

322 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

where c0 (t) it the N × 1 channel vector, s0 (t) is the useful signal transmitted
by the desired user, and n(t) is an additive white Gaussian noise vector.
Under optimal operation, the output of the combiner is given by

y = w0H x, (5.50)
where w0 is the optimal weight vector and the dependence on time is under-
stood.
Therefore, the ideal SNR at the output of an MRC combiner is
34 4 5
4wH c0 s0 42
0
γ0 = 34 4 5 , (5.51)
4wH n42
0

where )·* denotes expectation and | · | denotes magnitude. Reducing the expec-
tations, the above becomes:

|w0H c0 |2 σs2
γ0 = , (5.52)
w0H w0 σn2
where σs2 is the average transmitted power of the desired signal and σn2 is the
average noise power.
Substitution of the optimal weight vector yields
/2
σs2 cH
.
0 c0 σs2 cH
0 c0
γ0 = H
= . (5.53)
2
σn c0 c0 σn2
In a Rayleigh fading environment, the spatial signature c0 is an N -element
complex central Gaussian vector with covariance matrix R0 . If the fading is
uncorrelated then
R0 = σf2 IN , (5.54)
where σf2 is the variance of the fading process.
Then, the optimal SNR can be expressed
N
σs2 (
γ0 = 2 |c0 [n]|2 , (5.55)
σn n=1

where the terms of the sum are i.i.d 2 degrees-of-freedom chi-square variates
(exponential variates).
It follows that the ideal signal to noise ratio γ0 is then a chi-square variate
with 2N degrees of freedom:
) *N −1 γ0
γ0
1 γ̄c e− γ̄c
fγ0 (γ0 ) = , (5.56)
γ̄c Γ(N )
Diversity Combining 323

σs2 σf2
where γ̄c = σn 2 .

MRC with the effect of channel estimation

A significant portion of the literature on maximal-ratio combining often as-
sumes coherent reception on each branch, i.e. that each branch is equipped
with means to accurately estimate not only the channel gain, but also the car-
rier phase. Estimation of the latter can be performed using a standard carrier
recovery loop (e.g. Costas loop) or by relying on a pilot tone or signal. Such
an architecture is illustrated in Figure 5.7 where it is assumed that the com-
plex weights w1 , . . . wN are estimated with enough precision and updated of-
ten enough to keep track of changes in the carrier phase. The weight w(n) is
here split into its real and imaginary parts and QPSK / QAM modulation is
assumed.
Since MRC architectures in the spirit of Figure 5.7 require truly coherent
reception (where the coherence is provided by the weight estimation mecha-
nism), it has been stated that differential modulation schemes are irrelevant in
this context [Stuber, 1996]. However, several authors have derived MRC per-
formance results, e.g. with DPSK modulation [Proakis, 2000], [Zhang, 1998],
[Tomiuk, 1999]. To clarify this matter, the following aims to show that it
is possible to devise simpler MRC architectures which perform co-phasing
among branches without extracting an explicit phase reference, thus yielding
the same complexity advantage that single branch non-coherent receivers have
over their non-coherent counterparts.
It may seem curious, given the fact that some kind of phase estimate is re-
quired on each branch to compute the weight vector w, that we do not employ
it as a carrier phase estimate for coherent demodulation. This is justifiable
because the phase estimates for the purpose of weight generation (which are
matched to the vector channel) require a lot less precision than a carrier phase
estimate used for coherent demodulation. For example, a phase deviation in the
weights corresponding to ±1% of the symbol duration will cause an insignif-
icant SNR penalty at the receiver output. Carrier recovery, however, is much
more critical because of the higher frequency of the carrier [Proakis, 2000].
Indeed, a small fraction of the symbol period can correspond to a full or par-
tial phase reversal, thus leading to catastrophic performance. Furthermore, a
carrier phase error will introduce crosstalk between the in-line and quadrature
rails of QPSK and QAM modulation schemes.
Another factor to take into account is that typically, the physical channel
will vary relatively slowly with respect to the bit rate. Thus, the training pro-
cedure need only be performed once in an interval smaller than — but of the
order of — the coherence time of the channel. On the other hand, phase noise
in the local oscillator at the receiver and / or in the carrier of the received sig-
324 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

nal can vary very rapidly and requires closer tracking, such as provided by
conventional carrier recovery schemes.
These observations lead to MRC architectures such as depicted in Figure
5.8. Figure 5.8a shows an architecture where combining is performed at base-
band and prior to detection. The weights w(1) . . . w(N ) are computed using
the training sequence and vary relatively slowly. “Coherent” reception is in-
sured by weights pi , pr which are applied to the combined signals and track
the phase error introduced in the baseband signal by carrier phase changes.
This can be performed in a number of ways, including correlation with the
training sequence during the training interval and decision-directed tracking
afterwards.
Another possibility is depicted in Figure 5.8b where combining is performed
at the IF level. The weights are applied through adjustable delay elements or
phase shifters. A single carrier recovery loop is then required to bring the com-
bined signal to baseband prior to matched filtering. The weight estimation pro-
cedure (not shown) could involve upconverting the training sequence to IF so
that it can be correlated with branch signals using mixers and passive integra-
tors. It is also conceivable, depending on various practical system parameters
(cost of components, physical size constraints, electromagnetic compatibility,
wavelength and bandwidth), to perform the combining at RF. Compared with
the receiver of Figure 5.8a, this would bring RF hardware complexity down by
trading N downconverters for 1 downconverter and 1 upconverter.
Having established that carrier phase tracking is more costly than channel
tracking, it follows that the main advantage of the receivers in Figure 5.8 is that
a single carrier phase tracking device is required (operating on the combined
signal) instead of N . Furthermore, the carrier recovery loop in Figure 5.8b
could be replaced by a simpler non-coherent detector if DPSK is employed.
Yet another possibility is to have a postdetection combining scheme as pic-
tured in Figure 2 of [Zhang, 1998] where a DPSK detection chain is employed
in every branch. In such a scheme, the carrier reference in every branch is ob-
tained in a differential fashion, i.e. is provided by correlation with the previous
symbol, thus completely eliminating the need for any absolute phase reference.
Let us now evaluate the impact of desired signature estimation with respect
to this ideal SNR. The estimate is computed with the help of a known training
sequence s[k] which is K symbols long and is given by

K K
( |s[k]|2 c0 ( s∗ [k]n[k]
ĉ0 = + , (5.57)
K K
k=1 k=1

where the first term is the useful part of the estimate while the second term is
the estimation error or noise.
Diversity Combining 325

We can now compute the signal-to-estimation-noise ratio (SENR) of the

estimate ĉ0 :
) 6K *2
1 2
K k=1 |s[k]| cH0 c0
γ (ĉ0 ) = 7) *2 8 , (5.58)
1 6K ∗
K k=1 s [k]n[k]

which can be simplified as follows:

K 2 σs4 cH
0 c0 Kσs2 cH
0 c0
γ (ĉ0 ) = 2 2
= . (5.59)
KN σs σn N σn2
Therefore, the estimated signature can be expressed as the sum of the real
signature and an error vector:

ĉ0 = c0 + e, (5.60)
where, from (5.59), e is a spatially white Gaussian vector with covariance
matrix

σn2
Ree = IN = sIN , (5.61)
Kσs2
2
σn
and s = Kσ 2 is the average estimation noise-to-signal ratio per branch. Hence,
s
it can be verified that

cH0 c0 cH
0 c0
γ (ĉ0 ) = = , (5.62)
)eH e* Ns
which agrees with (5.59).
It is convenient to divide the error vector into two components with one
having the same direction in N -dimensional space as the real signature and the
other being orthogonal to it. This can be expressed mathematically by applying
to ĉ0 a unitary transformation P which brings the real signature entirely in the
first dimension.
Therefore, we have:

d̂0 = d0 + f , (5.63)
where
9
d0 = Pc0 = cH
0 c0 u, u = [1, 0, · · · , 0]T , (5.64)

d̂0 = Pĉ0 , (5.65)

326 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

and f is a Gaussian vector with covariance matrix Ree (since e is spherically

distributed and therefore invariant to rotations).
Separating out the in-line and orthogonal components of f , we get:
9
d̂0 = (1 + g(1))d0 + h cH
0 c0 , (5.66)
f
where g(1) is the first element of g = √ and h = [0, g(2), · · · , g(N )] is
cH
0 c0
the orthogonal component of g.
It follows that the SNR at the output of the combiner can be expressed as:

σ 2 |ĉH c0 |2 σ 2 |d̂H d0 |2
γ = 3.s 0 /2 5 = 7)s 0 *2 8 , (5.67)
ĉH0 n d̂H0 m

where m = Pn is the rotated noise vector. 9

Substituting (5.66) in the above and noting that d0 (1) = cH
0 c0 , we get:
/2 2
|1 + g(1)|2 cH
.
0 c0 σs
γ=. / . (5.68)
|1 + g(1) | c0 c0 + h hcH
2 H H 2
0 c0 σn

It follows that the normalized SNR ρ can be expressed as:

γ γσ 2 1
ρ= = H n2 = hH h
. (5.69)
γ0 c0 c0 σs 1 + |1+g(1)| 2

It follows that a = hH h is a scaled 2N − 2 degrees-of-freedom central

chi-square and its distribution is [Proakis, 2000]:
. aγ0 /N −2 aγ0
1 s e− s
fa (a) = . (5.70)
s Γ(N − 1)
Likewise, b = |1 + g(1)|2 is a scaled non-central, 2 degrees-of-freedom
chi-square variable with distribution [Proakis, 2000]:

γ0
) zγ * γ e− bγs0
− 0 0
fb (b) = e s
0 F1 1; . (5.71)
s s
It follows that c = ab is a ratio of chi-square variables. Hence, c is charac-
terized by a non-central F -distribution [Muirhead, 1982]:
γ0
(N − 1)e−
+ ,
s cγ0
fc (c) = 1 1 N, 1;
F . (5.72)
(1 + c)N s(1 + c)
Diversity Combining 327
1
Since ρ = 1+ γ1 c
, the pdf of ρ conditionned on γ0 can be found by applying
0
successive simple transformations on the random variable c (i.e., y = 1/x,
y = 1 + x/γ0 , y = 1/x) to finally obtain:

γ0
) ργ0 *
fρ|γ0 (ρ|γ0 ) = (N − 1)e− s (1 − ρ)N −2 1 F1 N, 1; . (5.73)
s
which is in fact a type of non-central beta distribution. Figure 5.9 shows the
distribution plotted for N = 10, 4 and 2 with s = 0.4 (i.e. could signify
2
that K = 5 and that σσ2s = 1/2) and γ0 = 1. For N = 4, a histogram is
n
superimposed which was obtained by running a 500000 iteration Monte-Carlo
simulation of a maximal-ratio combiner in a Rayleigh fading environment with
perfect power control.

Non-fading environments
If γ0 is considered fixed, the distribution of ρ is effectively given by (5.73).
This result is useful in environments which undergo no fading or which un-
dergo slow fading and it is desirable to study the performance over short peri-
ods of time (i.e. smaller than the coherence time of the vector channel). Non-
fading environments include free space, static communication systems with
little or no scatterer motion and communication systems with perfect power
control.
Knowing γ0 and the noise power σn2 , we can treat s as a design parameter
since it is a function of K, the training sequence length. To determine the
appropriate value for K to obtain on average a certain desired performance, it
is interesting to know the relationship between s and the average normalized
SNR ρ. This is readily obtainable in closed-form by simply integrating the
product of ρ and its distribution using [Gradshteyn and Ryzhik, 2000, 7.512-
12]:

# ∞
)ρ* = ρf (ρ) =
0
γ
− s0
e ) γ0 *
= 2 F2 2, N ; 1, 1 + N ; . (5.74)
N s
This function is plotted for various values of N and for γ0 = 1 in Figure
5.10. It can be seen that the curves drop very abruptly at first with increasing s
and then tend asymptotically towards N1 . This can easily be verified by taking
the limit of (5.73) when s → ∞. Two conclusions can be drawn from this
relation with the first being fairly obvious:
If the training sequence is very long (i.e. K → ∞), ρ tends towards 1.
328 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

If the training sequence is very short (such that s → ∞), the combiner be-
comes useless since the average performance is reduced to that of a single-
antenna receiver.
It is also of interest to find an expression for normalized SNR outage prob-
ability, i.e. the probability that ρ is smaller than some fixed threshold ρo . This
is found simply by integrating the PDF between 0 and ρo :

ρ0
ργ0 *
#
γ0
)
P (ρ < ρ0 ) = (N − 1)e− s (1 − ρ)N −2 1 F1 N, 1; dρ. (5.75)
0 s
This integral can be solved analytically by first expanding the factor (1 −
ρ)N −2 according to the binomial theorem. By interchanging the order of the
integration and summation operations, we are left with the simpler problem
of integrating the product of a power of ρ and the confluent hypergeometric
function. Applying [Gradshteyn and Ryzhik, 2000, 7.613-1] to each term, we
obtain:

N −2 + ,
−
γ0 ( N −1
Po (ρ0 ) = P (ρ < ρ0 ) = e s (−1)k ρk+1
o ×
k+1
k=0
) ρo γ0 *
1 F1 N ; k + 2; . (5.76)
s
Figure 5.11 shows plots of the normalized SNR outage probability versus s
when the outage point is ρ0 = 0.5 for various values of N . There is a narrow
region between s = 0 and s = 0.1 where the outage probability is null which
widens as N increases. Furthermore, the curves seem to tend asymptotically
with s to ever lower values as N increases.
In fact, if we let s tend towards infinity, (5.75) reduces to a very simple
expression which is solely a function of ρ0 and N :

Po∞ (ρ0 ) = lim P (ρ < ρ0 ) = 1 − (1 − ρ0 )N −1 . (5.77)

s→∞
It can be seen clearly from the form of this expression that to obtain an
asymptotic outage probability that is significantly lower than 1, the outage
point and / or the number of antennas must be small.
This illustrates how fast the bias towards Po∞ = 1 decreases with N . The
main implication of this from a system design perspective is that a receiver
should be operated either with low values of s (i.e. long enough training se-
quence to situate performance far enough away from the asymptotic region) or
be designed so that a low outage point is acceptable. Furthermore, for all other
parameters being equal, the training sequence length K should be proportional
to N to maintain a similar performance accross different array sizes.
Diversity Combining 329

Rayleigh fading environments

In uncorrelated Rayleigh fading, the vector channel c0 is an N -element
spherically symmetric Gaussian random vector, which implies that the opti-
mum SNR follows a 2N degrees-of-freedom chi-square distribution given by
(5.56).
Using the said distribution, we can obtain the distribution of ρ by performing
the following integral:
# ∞
fρ (ρ) = fρ|γ0 (ρ|γ0 )fγ0 (γ0 )dγ0 , (5.78)
0

which has a closed-form solution by virtue of [Prudnikov et al., 1992, p. 510,

eq. 2]:
- 0
(N − 1)(1 − ρ)N −2 ρ
fρ (ρ) = 2 F1 N, N ; 1; . (5.79)
1 + γ̄sc
. γ̄c /N
s +1

In this case, the average normalized SNR )ρ* is found by virtue of [Prud-
nikov et al., 1990, 2.21.1-4 and 7.4.1-3] to be:
- 0
1 1
)ρ* = /N 2 F1 1, 1; N + 1; + (5.80)
N γ̄sc + 1 1 + γ̄sc
.
- 0
N 1
* 2 F1 1, 1; N + 2; .
1 + γ̄sc
)
(N + 1) 1 + s γ̄c

Also, we can use the same term-by-term integration technique employed in

section IV (in obtaining (5.75)) and [Prudnikov et al., 1990, 2.21.1-4] to find
the normalized SNR outage probability which is expressed as:
N −2 +
(−1)k ρk+1
,
( N −1 0
Po (ρ0 ) = ×
k+1 .
1+ sγ̄c /N
k=0
- 0
ρ0
3 F2 k + 1, N, N ; 1, k + 2; ,
1 + γ̄sc

which also tends asymptotically to (5.77), although not at the same speed.
Figure 5.12 plots the outage probability for an outage point of ρo = 0.5
versus s for values of N ranging from 2 to 16. It is interesting to note that there
exists a trend reversal (in the vicinity of s = 0.6) in the ordering of the curves
with increasing N . This reflects the fact that a longer training sequence (and
hence, a lower s) is required when the array size N is augmented to maintain
330 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

a given performance level. As we let the training sequence length tend toward
zero, its impact becomes irrelevant and the ordering trend is reversed.
Finally, the PDF of the overall SNR γ, which includes both the effects of
multipath fading and of imperfect signature estimation, can be obtained by
evaluating [Garcia, 1994]
# ∞
1 )γ *
fγ (γ) = fρ,γ , x dx
γ |x| x
# ∞
1 ) γ4 *
= fρ|γ0 4 x fγ0 (x)dx, (5.81)
4
γ x x

which yields

(N − 1)1 F1 N, 1; γs
. /# ∞ ! "
1 1
N −2 −x s + γ̄c
fγ (γ) = (x − γ) e dγ,
Γ(N )γ̄cN γ
! "
1
−γ + γ̄1
e s c ) γ*
= /N −1 1 F1 N, 1; . (5.82)
s
. γ̄c
γ̄c s +1

Alternatively, this PDF can be expressed as a finite sum of simple terms (see
Appendix A for necessary manipulations) as follows:

− γ̄γ N −1 + , . γ /k
e c ( N −1 s
fγ (γ) = . (5.83)
γ̄c
. γ̄c
+1
/N −1 k k!
s k=0

It is assumed here that the combiner has the architecture of Figure 5.8, with
the coherent detection circuitry replaced by differential detection circuitry op-
erating on the combined signal. Given the expression for the bit error proba-
bility (BER) in noncoherent DPSK:
1
PbDPSK = e−γ , (5.84)
2
the corresponding PDF can be obtained by a simple transformation of random
variables: ) *
−ln(2Pb )
21 F1 N, 1; s
fPb (Pb ) = 1 1 /N −1 . (5.85)
(2Pb ) s + γ̄c γ̄c γ̄sc + 1
.

Likewise, the average bit error probability can be obtained by solving

! "
∞ −γ 1+ 1s + γ̄1
e γ*
# c )
P2DPSK = F
/N −1 1 1 N, 1; dγ, (5.86)
s
. γ̄c
0 2γ̄c s +1
Diversity Combining 331

which yields, by virtue of [Prudnikov et al., 1992, 3.35.1-1],

) *N −1
1 + 1s + 1
γ̄c
P2DPSK = ) *N . /N −1 . (5.87)
1 γ̄c
1+ γ̄c s +1 γ̄c

This is actually a simplified form of a result obtained by Tomiuk [Tomiuk,

1999], once the parameters s and γc are expressed in terms of Tomiuk’s and
Gans’ correlation parameter (see Appendix 5C).
Figure 5.13 shows the average error probability P2DPSK as a function of the
2
average bit SNR γ̄b = γ̄2c with s = 0.2 (i.e. could mean K = 5, σσs2 = 1 or
n
2
K = 3, σσ2s = 1.67) for values of N ranging from 1 to 12. It is seen therein
n
that augmenting N lowers P2 significantly even when the training sequence is
not lenghtened proportionally.
Figure 5.14 shows the average error probability in non-coherent DPSK re-
ception as a function of γ̄c with N = 8 for various values of K. The loss due
to training is seen to be relatively small as long as K is of the same order as
N (i.e. K ≥ 5 in the figure). A training loss of roughly 1.5 dB is observed
when K = 1 while for K = 0, performance is reduced to that of a one-antenna
receiver.
Using the finite-sum PDF from (5.83), simple term-by-term integration yields
the outage probability in Rayleigh fading:

N −1 + n
(γout /γ̄c )k
,
−
γout ( N −1 (
Po (γout ) = 1 − e γ̄c γ̄cn sN −1−n , (5.88)
n k!
n=0 k=0

which is equivalent to a result by Gans [Gans, 1971].

In many cases, the outage probability is considered a better performance
index of wireless communication systems than the average error probability.
Indeed, the average error probability by itself provides no information on the
magnitude of the performance fluctuations. For example, the penalty incurred
for insufficient training seems hardly significant in Figure 5.14 while it is more
obvious in Figure 5.15 which plots the outage probability as a function of the
average SNR γ̄c with γout = 1 for various values of K. The observed training
loss for K = 1 is now much more severe at about 6.5 dB.
What follows assumes a combiner architecture equivalent to Figure 5.8a or
Figure 5.8b with perfect carrier phase estimation. This differs from Proakis’
results for imperfect estimation with MPSK modulation which assume an ar-
chitecture like Figure 5.7 [Proakis, 2000], [Proakis, 1968].
Given the expression for the bit error probability for coherent QPSK:
332 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING
+1 ,
1 γ
PbQP SK = erfc , (5.89)
2 2
it is also possible, using the result (5.83) and integration by parts (see Appendix
5D), to obtain a closed-form expression for the average error probability:

N −1 +
: , 1
1 N −1 γ̄c
P2QP SK
(
= 1− γ̄cn sN −1−n ×
2 n γ̄ c + 2
n=0
n + ,+ ,k ;
( 2k 1
, (5.90)
k 2γ̄c + 4
k=0

which is equivalent to a result obtained by Tomiuk [Tomiuk, 1995]. Further-

more, letting s tend towards infinity to approach perfect training, the above
reduces to

n +
: 1 ,+ ,k ;
1 γ̄c ( 2k 1
P2QP SK = 1− , (5.91)
2 γ̄c + 2 k 2γ̄c + 4
k=0

which is equivalent to a result found in [Proakis, 2000] and [Proakis, 1968].

Similar results could in principle be obtained for other modulation schemes
(e.g. QAM) using the same techniques. However, multilevel modulation intro-
duces the additional impact of imperfect decision threshold estimation [Wil-
son, 1999] and this would have to be incorporated in the analysis.

5. Equal-gain combining
Since this scheme simply sums up the outputs of all branches without indi-
vidual weighting, it does away with the requirement for weight vector train-
ing and tracking. However, channel phase estimation is still required in each
branch to perform co-phasing prior to combining. It has been shown that for
small arrays, the performance of equal-gain combining was reasonably close to
that of MRC in fading, interference-free environments. In fact, in the limiting
case of an infinite number of branches, the difference in average SNR between
MRC and EGC is only 1.05 dB.

APPENDIX 5.A: Solution of a BER integral for selection di-

versity
Given the fact that for a basic selection diversity combiner operating in a Rayleigh fading
environment, the BER of noncoherent DPSK is given by
! ∞
N
! "
− γ −γ γ1 +1
P2DP SK = (1 − e γ̄c )N−1 e ¯c dγ, (5.A.1)
2γ̄c 0
APPENDIX 5.B: Conversion of fγ (γ) to finite sum form 333

we wish to solve an integral of the form

! ∞
N
! "
− γ̄γ −γ γ1 +A
I= (1 − e c )N−1 e ¯c dγ. (5.A.2)
2γ̄c 0

First, we perform a binomial expansion of the N − 1 order polynomial to obtain

N−1 # $! ∞ %
N " N −1 − γ
&k+1
I = e γ̄c (−1)k e−Aγ dγ,
2γ̄c k 0
k=0
N−1 # $ ! ∞
N " N −1
! "
−γ k+1 +A
= (−1)k e γ̄c dγ, (5.A.3)
γ̄c k 0
k=0

where the remaining integral is readily solvable and yields

 ! " ∞
N−1 # $ −γ k+1 +A
N " N −1 k  −e
γ̄c
I = (−1) 
2γ̄c k=1 k k+1
+A
γ̄ c
0
N−1
"# $ k
N N −1 (−1)
= . (5.A.4)
γ̄c k k+1
+A
k=1 γ̄c

We note that, by virtue of the Gamma function reflection formula (2.197),

(N − 1)! Γ(1 − N + k)
= (−1)k . (5.A.5)
(N − 1 − k)! Γ(1 − N )
We can also express the denominator polynomial in k (eq. (5.A.3)) as a ratio of Gamma
functions as follows:

1 γ̄c Γ (k + 1 + Aγ̄c )
k+1
= . (5.A.6)
γ̄c
+A Γ (k + 2 + Aγ̄c )

Substituting (5.A.5) and (5.A.6) in (5.A.3), we get

N−1 + ,
N " 1 1 − N + k, k + 1 + Aγ̄c
I= Γ . (5.A.7)
2 k=1 k! 1 − N, k + 2 + Aγ̄c

We note that in the above, the summation can be extended from k = 0 to k = ∞ since
Γ [1 − N + k1 − N ] = 0 for all k > N − 1. Therefore, we can express this summation as a
Gauss hypergeometric function:
+ , # - $
N 1 + Aγ̄c 1 − N, 1 + Aγ̄c --
I= Γ 2 F1 1 , (5.A.8)
2 2 + Aγ̄c 2 + Aγ̄c -
which, by virtue of the unity argument and identity [Prudnikov et al., 1990, 7.3.5-1], reduces to
+ , + ,
N 2 + Aγ̄c , N 1 1 + Aγ̄c , N + 1
I= Γ = Γ . (5.A.9)
2 (1 + Aγ̄c ) 1 + Aγ̄c + N 2 1 + Aγ̄c + N
334 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

APPENDIX 5.B: Conversion of fγ (γ) to finite sum form

Application of the Kummer identity [Prudnikov et al., 1990, 7.2.2-8] to the confluent hyper-
geometric function in the following:
! "
−γ 1 + γ̄1
e s c
% γ&
fγ (γ) = /N−1 1 F1 N, 1; , (5.B.1)
s
. γ̄c
γ̄c s
+1
will yield an exponential and a confluent hypergeometric function with a negative first argument:
! "
−γ 1 + γ̄1
e s c γ
% γ&
fγ (γ) = /N−1 e s 1 F1 1 − N, 1; − . (5.B.2)
s
. γ̄c
γ̄c s
+1
The latter is in fact a hypergeometric polynomial [Prudnikov et al., 1990] expressible as a
finite sum:

− γ̄γ N−1
e c " Γ(1 − N + k) % γ &k
fγ (γ) = /N−1 − . (5.B.3)
Γ(1 − N )k! s
. γ̄c
γ̄c s
+1 k=0

Applying the reflection formula Γ(z−k)

Γ(z)
Γ(1−z)
= (−1)k Γ(1−z+k) , the ratio of negative-argument
Gamma functions can be changed to positive argument Gamma functions or factorials, thus
yielding

− γ̄γ N−1
"# $% &
e c N −1 γ k
fγ (γ) = . (5.B.4)
γ̄c
. γ̄c
+1
/N−1 k s
s k=0

APPENDIX 5.C: Equivalence with Gans’ result

The paper by Gans [Gans, 1971] considers that the weights employed to combine an un-
correlated Rayleigh-fading vector channel are Gaussian random variables pk . Since they are
obtained through imperfect estimation, they are not equal to, but are correlated with, the Gaus-
sian random variables gk describing the real channel. His results are expressed in terms of the
correlation parameter ρ (not to be confused with our normalized SNR ρ) which is defined as
follows:

- #pk gk∗ $ -2
- -
ρ2 = -- - . (5.C.1)
#pk p∗k $ -
Furthermore, Gans imposes, without loss of generality2 , that pk and gk have the same vari-
ance.
Let s0 c0 be, in our notation, the vector of channel coefficients gk as defined by Gans. Fur-
thermore, let the vector p be the vector of imperfect weights pk , adjusted to have the same
variance as s0 c0 . Therefore, we have

(s0 c0 + e)σs2 #cH

0
0 c0 $
p= 0 . (5.C.2)
H
#(s0 c0 + e) (s0 c0 + e)$
It follows that

2 Indeed, a gain applied to all branches does not affect the output SNR.
 QP SK
APPENDIX 5.D: Derivation of P2 335
2
1 H 2
2 1 |#(s0 c0 + e)s0 c0 $| c0 c0
|#pk gk∗ $| = 2 , (5.C.3)
N #(s0 c0 + e)H (s0 c0 + e)$
and
2
1 H 22 1 H 22
2 1 |#(s0 c0 + e)(s0 c0 + e)$| c0 c0 c0 c0
|#pk p∗k $|
= 2 2
= . (5.C.4)
N H
#(s0 c0 + e) (s0 c0 + e)$ N2
Hence, from (5.C.1), (5.C.3) and (5.C.4), we have

|#(s0 c0 + e)s0 c0 $|2 cH σs2 cH

1 2 1 2
2 0 c0 0 c0
ρ = 2
= 2 H
#(s0 c0 + e)H (s0 c0 + e)$ #cH
0 c0 $
σs #c0 c0 $ + #eH e$
γ̄c σs2
= , (5.C.5)
γ̄c σs2 + σn2 /K
2
σn
which, noting that s = Kσs2, reduces to
γ̄c 1
ρ2 = = . (5.C.6)
γ̄c + s 1 + γ̄sc
It follows that
s 1
1 − ρ2 = = , (5.C.7)
γ̄c + s 1 + γ̄sc
and

ρ2 γ̄c
= . (5.C.8)
1 − ρ2 s
Substituting (5.C.6), (5.C.7) and (5.C.8) in (5.B.4), we get
− γ̄γ N−1 # $k
(1 − ρ2 )N−1 " ρ2 γ
$#
e c N −1
fγ (γ) = , (5.C.9)
γ̄c k (1 − ρ2 )γ̄c
k=0

which is exactly the output SNR pdf obtained by Gans.

APPENDIX 5.D: Derivation of P2QP SK

The average error probability in QPSK can be expressed as follows
γ
. /
1 F1 N, 1; s
! ∞ #3 $
γ
! "
1+ 1
P2QP SK =
−γ s
/N−1 e γ̄c erf c dγ. (5.D.1)
2
. γ̄c
0 +1 γ̄c
s
Integration-by-parts constitues an appealing strategy here since the derivative of the erfc(·)
γ
.0 γ / −
e 2
factor has a very simple form. Letting u = erfc 2
, we have du = − √ 2πγ
.
Accordingly,
x
! γ . /
1 F1 N, 1; s
! "
−x 1 + γ̄1
v= . γ̄c /N−1 e s c dx, (5.D.2)
0
s
+1 2γ̄c
which is easily solved by substituting our result for the outage probability from (5.88), i.e.
v = Po2(γ) .
Applying integration by parts and noting that the uv term between 0 and ∞ is null, we obtain
336 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

4 3 N−1 # $
1 2γ̄c " N −1
P2QP SK = 1− γ̄cn sN−1−n ×
2 γ̄c + 2 n=0 n
# % &−1 $k 
1 1 1
n γ̄c 2
+ γ̄c # $
" 1
√ Γ + k , (5.D.3)

k=0
2πk! 2 

where

N−1
"# $
N −1
γ̄cn sN−1−n =
n
n=0
N−1
"# $# $−n %
N −1 s γ̄c &n+1−N
1+ 1+ = 1, (5.D.4)
n γ̄c s
n=0

by virtue of the binomial theorem. Hence, we have

N−1 # $
1 " N −1
P2QP SK = γ̄cn sN−1−n ×
2 n=0 n
 % &k . /
2 1
3
γ̄c "n
γ̄c +2
Γ 2
+ k
1 − √ , (5.D.5)
 
γ̄c + 2 k=0 πk!

√
π22k Γ(k + 1)Γ k + 12 (see Appendix
. /
which, after application of the identity Γ(2k + 1) =
II of [Prudnikov et al., 1986b]), becomes

N−1 # $
1 " N −1
P2QP SK = γ̄cn sN−1−n ×
2 n=0 n
4 3 n # $# $k 7
γ̄c " 2k 1
1− . (5.D.6)
γ̄c + 2 k 2γ̄c + 4
k=0

References

[Proakis, 2000] J. G. Proakis, Digital Communications, 4th ed.. New York: McGraw-Hill.
[Jakes, 1974] W. C. Jakes, “Multipath interference,” chapter in Microwave mobile communi-
cations. William C. Jakes, ed., New York: J. Wiley & Sons.
[Pierce, 1958] J. N. Pierce, “Theoretical diversity improvement in frequency-shift keying,”
Proc. IRE, vol. 46, pp. 903Ð910, May 1958.
[Ritcey and Azizo, 1998] J. A. Ritcey and M. Azizog̃lu, “Impact of switching constraints on
selection diversity performance,” in Proc. 32nd Asilomar Signals, Systems and Computer
Conf., vol. 1, pp. 795-9, Pacific Grove, California, Nov. 1998.
REFERENCES 337

[Barnard and Pauw, 1989] J. H. Barnard and C. K. Pauw, “Probability of error for selection
diversity as a function of dwell time,” IEEE Trans. Commun., vol. 37, no. 8, pp. 800-3,
Aug. 1989.
[Neasmith and Beaulieu, 1998] E. A. Neasmith and N. C. Beaulieu, “New results on selection
diversity,” IEEE Trans. Commun., vol. 46, no. 5, pp. 695-704, May 1998.
[Wang et al., 2001] Q. Wang, M. Lecours and C. Vergnolle, “Criteria for wide-band radial
switch design,” IEEE Trans. Microwave Theory and Techniques, vol. 49, no. 1, Jan. 2001.
[Kim and Kim, 2001] Y. G. Kim and S. W. Kim, “Optimum selection diversity for BPSK sig-
nals in Rayleigh fading channels,” IEEE Trans. Commun., vol. 49, no. 10, pp. 1715-8, Oct.
2001.
[Pozar, 1998] David M. Pozar, Microwave engineering, 2nd ed. New York: J. Wiley & Sons,
1998.
[Muirhead, 1982] R. J. Muirhead, Aspects of Multivariate Statistical Theory. New York: J.
Wiley & Sons, 1982.
[Prudnikov et al., 1990] A. P. Prudnikov, Y. A. Brychkov and O. I. Marichev, Integrals and Se-
ries vol. 3: More Special Functions. Amsterdam: Gordon and Breach Science Publishers,
1990.
[Stuber, 1996] G. L. Stuber, Principles of Mobile Communications, Kluwer Academic Pub-
lishers, Boston, 1996.
[Zhang, 1998] Q. T. Zhang, “Exact analysis of postdetection combining for DPSK and NFSK
system over arbitrarily correlated Nakagami channels," IEEE Trans. Comm., vol. 46, pp.
1459-1467, Nov. 1998.
[Tomiuk, 1999] B. R. Tomiuk, N. C. Beaulieu and A. A. Abu-Dayya, “General forms for max-
imal ratio diversitywith weighting errors,” IEEE Trans. Commun., vol. 47, April 1999, pp.
488-492.
[Gradshteyn and Ryzhik, 2000] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series
and Products, Academic Press, San Diego, 2000.
[Prudnikov et al., 1992] A. P. Prudnikov, Y. U. Brychkov and O. I. Marichev, Integrals and
Series, vol. 4: Direct Laplace Transforms, Amsterdam: Gordon and Breach.
[Garcia, 1994] A. Leon-Garcia, Probability and Random Processes for Electrical Engineering,
Addison-Wesly, Reading, 1994.
[Gans, 1971] M. J. Gans, “The effect of Gaussian error in maximal ratio combiners,” IEEE
Trans. Commun. Tech., vol. 19. Aug. 1971, pp. 492-500.
[Proakis, 1968] J. Proakis, “Probabilities of error for adaptive reception of M -phase signals,"
IEEE Trans. Commun. Tech., vol. COM-16, pp. 71-81, Feb. 1968.
[Tomiuk, 1995] B. R. Tomiuk, Effects of Weighting Errors in Maximal Ratio Combiners on
the Bit Error Rates of Digital Diversity Systems, M. Sc. dissertation, Queen’s University,
Kingston, Canada, 1995.
[Wilson, 1999] S. K. Wilson and J. M. Cioffi, “Probability density functions for analyzing
multi-amplitude constellations in Rayleigh and Ricean channels,” IEEE Trans. Commun.,
vol. COM-47, pp. 380-386, March 1999.
[Prudnikov et al., 1986b] A. P. Prudnikov, Y. U. Brychkov and O. I. Marichev, Integrals and
Series, vol. 2: Special Functions, Amsterdam: Gordon and Breach.
338 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

Figure 5.1. Proposed RF front-end architecture.

REFERENCES 339





  



  

Figure 5.2. Envelope detector circuit example.

340 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

Figure 5.3. A variant of the proposed RF front-end architecture.

REFERENCES 341

0.05 N =2
N =4
N = 10
0.04

0.03
fγ (γ)

0.02

0.01

0
0 10 20 30 40 50
γ

Figure 5.4. Distribution fγ (γ) for various array sizes when the average SNR σs2 /σν2 is 10 dB
(solid line = SDC1, dashed line = SDC2).
342 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

N =2
100
N =4

80 N = 10

N = 25
60
γ̄

40

20

0
0 5 10 15 20 25 30
σs2 /σν2

Figure 5.5. Average output SNR as a function of input SNR σs2 /σν2 (solid line = SDC1, dashed
line = SDC2).
REFERENCES 343

10−1

10−2
(QP SK,SD2)

N =2

N =4
10−3
, P2

N = 10
(QP SK,SD)

N = 25

10−4
P2

10−5

10−6 -2.5 0 2.5 5 7.5 10 12.5 15

σs2 /σν2

Figure 5.6. Error probabilities averaged over Rayleigh fading for QPSK-modulated signals as
a function of input SNR σs2 /σν2 (solid line = SDC1, dashed line = SDC2).
344 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

w(n)r
cos(2πfc t) from other
branches

LPF w(n)i in-line branch

+ yr zr
Matched
x(n) − Filter
+
Matched
+
yi Filter zi
LPF w(n)i
quadrature branch
from other
branches
sin(2πfc t)
w(n)r

Figure 5.7. An example of a classical predetection MRC receiver architecture.

REFERENCES 345

from other
w(n)r branches pr
cos(2πfc t)
in-line branch
+ yr zr
+ Matched
LPF w(n)i pi
− Filter
−

x(n)

+ +
w(n)i pi Matched
LPF +
+ yi Filter zi

quadrature branch
sin(2πfc t)
w(n)r from other pr
branches

(a)

from other in-line branch

w(n) branches yr zr
Matched
Filter
x(n) DC to IF
downconversion BPF ×
90◦ zr
Matched
yi Filter
from other
branches quadrature branch
Carrier
Recovery VCO

(b)

Figure 5.8. MRC receivers with separate channel and carrier tracking: (a) baseband-
combining predection MRC receiver with baseband compensation of carrier phase jitter; (b)
IF-combining predection MRC receiver with single carrier recovery loop; box marked × is
comprised of an adjustable delay element or phase shifter, followed by an adjustable attenuator.
346 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

3.5

3
N =2
2.5
N = 10
f (ρ)

2
N =4
1.5

0.5

0
0 0.2 0.4 0.6 0.8 1
ρ

Figure 5.9. Distribution of the normalized SNR ρ conditionned on γ0 when s = 0.4, γ0 = 1,

and the number of antennas N = 2, 4 and 10.

0.8

0.6
N =2
ρ

0.4

N =4
0.2
N =8
N = 16
0 1 2 3 4
s
γ0

Figure 5.10. Average normalized SNR ρ in non-fading environments as a function of s/γ0 for
N = 2, 4, 8 and 16.
REFERENCES 347

0.8
Po (ρo = 0.5)

0.6

0.4

0.2
N = 2, 3, 4, 5, 6, 8, 10, 12, 16

0
0.4 0.8 1.2 1.6 2
s
Figure 5.11. Outage probability of the normalized SNR ρ with no fading, γ0 = 1, ρo = 0.5
and for various values of N .

0.6

0.5
N = 2, 3, 4, 5, 6, 8, 10, 12, 16
Po (ρo = 0.5)

0.4

0.3

0.2

0.1

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
s
Figure 5.12. Outage probability of the normalized SNR ρ in Rayleigh fading with outage point
ρo = 0.5 and γ̄c = 1 for various values of N
348 SPACE-TIME METHODS, VOL. 1: SPACE-TIME PROCESSING

100
10−1
10−2
10−3
N =1

10−4
P2DP SK

10−5 N =2
10−6
10−7 N = 4, 6, 8, 10, 12

10−8

10−9

10−10 5 10 15 20 25 30 35
γ̄b (dB)

Figure 5.13. Average error probability P2 for non-coherent DPSK with s = 0.2 as a function
of γ̄b for various values of N .

100

10−1
10−2 K=0
10−3
10−4
P2DPSK

10−5 K = ∞, 20, 10, 5, 2, 1

10−6
10−7
10−8
10−9

10−10
5 10 15 20 25
γ̄c (dB)

Figure 5.14. Average error probability P2 for non-coherent DPSK with N = 8 for various
lengths K of the training sequence.
REFERENCES 349

100

K=0

10 −1

K = ∞, 20, 10, 5, 2, 1
Po

10−2

10−3
-5 0 5 10 15
γ̄c (dB)

Figure 5.15. Outage probability versus average SNR γ̄c for an outage threshold of γout = 1.